The disorder of information in any of its forms causes harm in multiple dimensions of life. In this regard, numerous efforts attempt to understand and detect fake news early. The use of machine learning models is very promising for the task of fake news detection through automatic classification. However, class imbalance seriously affects classification models, even state-of-the-art ones. We are particularly interested in works that address the problem by exploiting the text of fake news, using a hybrid approach: machine learning and natural language processing. There is an abundance of works, primarily in English, that tackle binary classification (often: real or fake), while fewer address the multiclass problem. We propose a methodology to improve the performance of information disorder detection models in text, specifically in imbalanced and multiclass datasets. Our approach is based on considering surface linguistic features related to the text, applying techniques for data balancing, and subsequently developing classification models. We experiment on the CLNews dataset, which contains rumors in Spanish and four imbalanced categories. Our experiments show how, by using our methodology, classification can be improved, raising the accuracy from 0.38 to 0.93 when using SMOTE to address the imbalance along with Random Forest for classification and from 0.35 to 0.93 when using ROS for imbalance along with XGBoost for classification.